Skip to content

Machine Learning (ML) - Notes

Src : towardsdatascience - Hunter Heidenreich

Agenda

Intro

ML is a field of AI algorithms that learn from EXAMPLES and EXPERIENCES instead of traditional HARDCODE and RULES

Ex : Apple and oranges detection

ML vs Traditional coding

  • Tradictional coding : too much rules

  • ML : Model / Classifier and train it to generate the RULES instead of writing them

  • classifier (as function) : takes a data as input (FEATURES) and signs LABEL to it as output(LABELS)

  • Fruit(apple or orange?) => Classifier => apple (if apple chose)
  • email (spam/mail_ok) => Classifier => spam (if mail_nok)

Training the Model (CLassifier)

To Train the Classifier we use :

  • Supervised Learning(SL) : it learns from examples/experiences
  • Unsuspervised Learning(USL) : it learns from events ?
  • Reinforcement Learning (RL): Conceptually similar to human learning processes
  • ex: a robot learning to walk
  • Strategy games : Go, Chess etc

The more the training data exists => the better the classifier Will be

ML train 

|--------|          |-----------|       |-----------|
|Collect |          |Train      |       |Make       |
|Training|    =>    |Classifier |   =>  |Predictions|
|   Data |          |           |       |           |
|--------|          |-----------|       |-----------|

ML algorithms

  • Supervised :
  • Regression : Predicting a continuous-valued attribute associated with an object
    • Multiple Linear Regression(MLR)
    • Polynomial Regression (PR)

  • Classification : Identifying which category an object belongs to

    • k-Nearest Neighbor
    • Decision Trees(ID3, C4.5, C5.0)
    • logistic regression
    • Naïve Bayes
    • Linear Discriminant Analysis
    • Neural Networks
    • Support Vector Machines (SVM)
    • Random Forest(RF)

  • Unsupervised :

    • Clustering : Automatic grouping of similar objects into sets.
      • k-Means
      • Mean-shift
      • Hierarchical Clustering (HC)
      • Density-based Clustering (DBSCAN)
      • Gaussian Mixture Models(GMM)
  • Reinforcement :
    • Q-Learning
    • Deep Q-Network (DQN)
    • A3C

ML Frameworks/tools

  • TensorFlow
  • PyTorch
  • Scikit-learn
  • Spark ML
  • Torch
  • Huggingface
  • Keras

ML Problem solving in 7 steps

  1. GATHERING / COLLECTING DATA : The more data we collect the more accurate will the model.

  2. We collect datas to train the model of the system we want to deploy

  3. DATA PREPARATION :

    • Features ? : the input of the system
    • Labels ? : the output of the system 
          ----------------------------------
    |       Features        | Labels|   
    ----------------------------------
    |   x1  |   x2  |   xn  |   y1  |
    ----------------------------------
    |   ..  |   ..  |   ..  |   ..  |
    ----------------------------------
    | xn,m |        |       |   yn,m|
    ----------------------------------
    • visualization of datas
    • balances, relationship between datas
    • split : training/evaluation(performance of the model)

  4. Choosing the MODEL : There are already alot of model created by DataScientists :

    • For : Music, image, number, text, text based data, linear model (y=ax+b)

  5. TRAINING (the model : Y = mx+b) :

    • Y : output
    • m : SLope (many m possible, as many features)
    • X : input
    • b : Y-intercept 
                  [m1,1 m1,2] 
    Weight = [m2,1 m2,2] 
            [m3,1 m3,2]

            [b1,1 b1,2] 
    biases = [b2,1 b2,2] 
            [b3,1 b3,2]
      Training process 
          |--------|          |-----------|       |-----------|
    |        |          |   Model   |       |           |
    |Training|    =>    |   (W,b)   |   =>  |Prediction |
    |   Data |          |           |       |           |
    |--------|          |-----------|       |-----------|

                        |-----------|           ||
        /\              |           |           \/
        ||      <=      |Test/update|   <=
                        | (W,b)     |   
                        |-----------|

Each iteration it's called, a training steps.

/!\ #Residus = Biais ( θ ^ ) ≡ E [ θ ^ ] − θ #Définition — Si θ ^ est l'estimateur de θ

  1. EVALUATION : after the model is good time to evaluate
        |----------|         |-----------|       |------------|
        |          |         |   Model   |       |            |
        |EVALUATION|  =>     |   (W,b)   |  =>   | Prediction |
        |   Data   |         |           |       |            |
        |----------|         |-----------|       |------------|

                            |-----------|           ||
            /\              |           |           \/
            ||      <=      |Test       |   <=
                            | (W,b)     |   
                            |-----------|

This metric allows the model to see the data that has not yet seen. This is to test how the model might act in the real world

  1. PARAMETER TUNING

    • To improve the training
    • Repeat the training data several time to increase the accuracy
    • Learning rate : limit of the train / how far we shift the line between two input datas
    • initial conditions : for complexes models (value = 0 ...)
    • Hyperparameters.

/!\ : it's important to choose the good parameters to be changed

  1. PREDICTIONS : ML uses datas to answer questions

Input : features Output : Labels

References

Scikit learn : https://scikit-learn.org/stable/#

TensorFlow : https://www.tensorflow.org/resources/learn-ml

Spark ML : https://spark.apache.org/docs/latest/ml-guide.html

PyTorch : https://pytorch.org/tutorials/beginner/deep_learning_60min_blitz.html https://docs.microsoft.com/en-us/learn/paths/pytorch-fundamentals/

Google course/ Josh gordon : https://www.youtube.com/watch?v=cKxRvEZd3Mw&list=PLOU2XLYxmsIIuiBfYad6rFYQU_jL2ryal

Google Cloud Plateform / Yufeng Guo: https://www.youtube.com/watch?v=nKW8Ndu7Mjw

IBM Cloud : https://www.ibm.com/cloud/blog/supervised-vs-unsupervised-learning